More
Best rig by budgetRent before you buySubmit a buildNewsletterFree datasetAdvertise
Submit build

Builds / RTX 3090 24 GB

RTX 3090 at 250 W: Qwen3.8 27B, llama.cpp 40.6 vs patched vLLM + MTP 129.8 tok/s

Last updated:

Source: alexander-ollman.github.io ↗📏 How measured: reported by the original source; RigBase re-checked the numbers at that source
GPU(s)
RTX 3090 24 GB
Amazon · eBay (used) · Newegg
VRAM / unified memory (GB)
24
CPU
not stated
Motherboard
not stated
RAM
not stated
PSU
not stated
Other
Test rig has two RTX 3090s; figures below are single-card unless noted. 250 W power cap, thinking disabled; driver 580.173; figures are the log's canonical single-card table (mean of 5 runs)
Reported total cost
not stated
Power
250 W cap

Measured speed (generation tokens/sec)

ModelQuantEnginetok/sPrompt processing (tok/s)Details
Qwen3.8 27BUD-Q4_K_XL (GGUF)llama.cpp, tuned40.6not stated⚡ spec-decode (MTP n=1)context: ~55-token prompt📏 Mean of 5; 42.6 at ~1K prompt; 27.6 with out-of-box defaults
Qwen3.8 27BAutoRound W4A16vLLM (stock)32.6not statedcontext: ~55-token prompt📏 Mean of 5, no speculation; 32.8 at ~1K prompt
Qwen3.8 27BAutoRound W4A16vLLM (patched)129.8not stated⚡ spec-decode (MTP k=4)context: ~55-token prompt📏 Mean of 5, acceptance 0.764; 167.8 at ~1K prompt. Stock vLLM + MTP k=4 did 66.5 / 85.6
Qwen3.8 27BUD-Q4_K_XL (GGUF)llama.cpp llama-bench, 1 GPU at 350 W38.81,027 @ 512-token prompt
est. TTFT @8k ≈ 8.0 s (calculated from measured prefill)		context: tg128📏 llama-bench, f16 KV, 350 W (q8_0 KV: 38.7 tg, 1,275 pp512)

Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 ÷ prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.

Public measurement log for Qwen3.8 27B on RTX 3090. Patched vLLM + DFlash2 k=7 reached 145.7 (~55 prompt) / 212.1 (~1K) tok/s. Two cards with llama.cpp layer split: 42.6 tg128 / 1,391 pp512 at 350 W (a second card buys +35% prefill, +10% decode). The log's Flash-Next run uses IQ3_S, below Q4, so it is not listed. Found via llamaperf (discovery only); numbers read at the original log on 2026-10-30.

Source: https://alexander-ollman.github.io/qwen3.8-on-rtx3090/log.html

Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.

Embed this result

Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds — numbers stay on the build page.

Markdown

[![Measured on RigBase](https://getrigbase.com/badge/rtx-3090-250w-qwen38-27b-vllm-mtp-vs-llamacpp-ollman.svg)](https://getrigbase.com/builds/rtx-3090-250w-qwen38-27b-vllm-mtp-vs-llamacpp-ollman/)

HTML

<a href="https://getrigbase.com/builds/rtx-3090-250w-qwen38-27b-vllm-mtp-vs-llamacpp-ollman/" rel="noopener"><img src="https://getrigbase.com/badge/rtx-3090-250w-qwen38-27b-vllm-mtp-vs-llamacpp-ollman.svg" alt="Measured on RigBase" height="20"/></a>

Preview: Measured on RigBase · How to use

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)

Quick answers: Can RTX 3090 run Qwen3.8 27B?